CN109241995A - A kind of image-recognizing method based on modified ArcFace loss function - Google Patents

A kind of image-recognizing method based on modified ArcFace loss function Download PDF

Info

Publication number
CN109241995A
CN109241995A CN201810866142.4A CN201810866142A CN109241995A CN 109241995 A CN109241995 A CN 109241995A CN 201810866142 A CN201810866142 A CN 201810866142A CN 109241995 A CN109241995 A CN 109241995A
Authority
CN
China
Prior art keywords
image
loss function
image recognition
modified
indicate
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201810866142.4A
Other languages
Chinese (zh)
Other versions
CN109241995B (en
Inventor
章东平
陈思瑶
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China Jiliang University
Original Assignee
China Jiliang University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by China Jiliang University filed Critical China Jiliang University
Priority to CN201810866142.4A priority Critical patent/CN109241995B/en
Publication of CN109241995A publication Critical patent/CN109241995A/en
Application granted granted Critical
Publication of CN109241995B publication Critical patent/CN109241995B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Abstract

The invention discloses a kind of image-recognizing methods based on modified ArcFace loss function, the feature of image is extracted using the image recognition network based on deep learning, the feature of extraction is used to describe the main information of image, the image recognition network based on deep learning is trained using modified ArcFace loss function, modified ArcFace loss function maximizes classification boundaries by not only reducing inter- object distance but also increasing between class distance in angular region, to improve the accuracy of image recognition model identification.The present invention is used for area of pattern recognition.

Description

A kind of image-recognizing method based on modified ArcFace loss function
Technical field
The invention belongs to the deep learning fields that deep neural network extracts characteristics of image, are related to neural network, mode Technologies more particularly to a kind of image-recognizing methods based on modified ArcFace loss function such as identification.
Background technique
With the arriving of big data era and greatly improving for computing capability, image recognition technology is just managed towards high-level semantics Solution direction is developed, and the image recognition technology based on deep learning has become the research hotspot of current artificial intelligence field.
Image recognition technology is that image is handled automatically, analyzed and understood by computer, to identify various differences The target of mode and the technology of object.Image recognition technology has a wide range of applications, for example, can pass through plant in agricultural Growing way, the color of leaf and flower judged, watered, applied fertilizer to plant, desinsection；In the industry, image can be passed through It identifies and control is carried out to the product in entire workshop；In medicine, the strong of analysis patient can be carried out by the shape of cell, bone Health situation；In space flight, space research can be carried out according to the Real-time Feedback of satellite photo；In daily life, image recognition Technology is also very universal, such as Car license recognition, fingerprint recognition；However, there is also some difficulties for image recognition technology, due to viewpoint Variation, the increase image recognition such as background is complicated, shadow change, blocks, deforming difficulty, cause in the image based on deep learning Identify image classification inaccuracy problem in network training process, to solve this problem, ArcFace loss function is suggested, but ArcFace only maximizes classification boundaries from reducing inter- object distance.
The key to solve the above problems is exactly to design a modified ArcFace loss function, based on deep learning In image recognition network training process, inter- object distance was not only reduced but also had increased between class distance to maximize classification boundaries, to improve The accuracy of image recognition model identification.
Summary of the invention
It is a kind of based on modified ArcFace loss function the present invention overcomes proposing in place of the deficiencies in the prior art Image-recognizing method, it is intended that carry out image recognition model training using modified ArcFace loss function, improve in reality scene The accuracy of image recognition.
The present invention is to adopt the following technical scheme that up to foregoing invention purpose
A kind of image-recognizing method based on modified ArcFace loss function, step include:
Step (1): prepare image recognition training dataset, test data set；
Step (2): image recognition network structure of the building based on convolutional neural networks, it is described based on convolutional neural networks Image recognition network includes convolutional layer, pond layer, full articulamentum, modified ArcFace loss function layer, wherein two convolution Layer constitutes an image recognition minor structure with a pond layer, and image recognition network is by N number of concatenated minor structure, two full connections Layer F1、F2, a modified ArcFace loss function layer composition；
Step (3): image recognition training dataset is input to the image based on convolutional neural networks of step (2) building Identify training in network, the loss function in training process uses modified ArcFace loss function, by constantly to network Carry out loop iteration training reduce loss function constantly, until complete setting the number of iterations Q, and by image recognition model into Row saves；
Further, the modified ArcFace loss function calculation formula are as follows:
Wherein, n indicates that the training sample sum of iteration input every time in training process, L indicate that the loss of n sample is average Value, yiIndicate the image category label of i-th of sample,Indicate full articulamentum F2Weight matrix yiColumn and full articulamentum F1Output angle, θjIndicate full articulamentum F2Weight matrix jth column and full articulamentum F1Output angle, s indicate Adaptive cosine coefficient (desirable s=64), C indicate the classification number of total training sample, λ indicate adaptive weighting coefficient (desirable λ= 0.5), m indicates decision edge, is obtained by network training；
Step (4): image recognition test data set is subjected to image spy using image recognition model obtained in step (3) Sign is extracted, and calculates the cosine similarity P between every two image feature vector, setting image similarity threshold value is T, if similar It spends P and is greater than threshold value T, then judge that two images are same class images, otherwise judge that two images are not same class images, obtain figure As the test result of identification model.
Compared with prior art, the beneficial effects of the present invention are embodied in:
The present invention takes a kind of modified ArcFace loss function that is based on to carry out image recognition, using based on deep learning Image recognition network the feature of image is extracted, the feature of extraction is used to describe the main information of image, using changing The image recognition network based on deep learning is trained into type ArcFace loss function, ArcFace loss function only exists Angular region maximizes classification boundaries by reducing inter- object distance, and different classes of image can not be made to divide as far as possible, And modified ArcFace loss function maximizes classification by both reducing inter- object distance increase between class distance in angular region Boundary can not only make same category of image more compact, also different classes of image can be made to divide as far as possible, to mention The accuracy of hi-vision identification model identification, the present invention are suitable for image recognition, can overcome using ArcFace loss function pair Inaccurate problem is identified based on the image recognition model that convolutional neural networks are trained, and improves the accurate of image recognition Property.
Detailed description of the invention
Fig. 1 is a kind of image recognition convolutional neural networks structural schematic diagram based on modified ArcFace loss function.
Specific embodiment
The present invention will be further explained below with reference to the attached drawings.
In the present embodiment, as shown in Figure 1, a kind of image-recognizing method based on modified ArcFace loss function includes Following steps:
Step (1): prepare image recognition training dataset, test data set；
Step (2): image recognition network structure of the building based on convolutional neural networks, it is described based on convolutional neural networks Image recognition network includes convolutional layer, pond layer, full articulamentum, modified ArcFace loss function layer, wherein two convolution Layer constitutes an image recognition minor structure with a pond layer, and image recognition network is connected entirely by 32 concatenated minor structures, two Meet a layer F1、F2, a modified ArcFace loss function layer composition；
Step (3): image recognition training dataset is input to the image based on convolutional neural networks of step (2) building Identify training in network, the loss function in training process uses modified ArcFace loss function, by constantly to network Carrying out loop iteration training reduces loss function constantly, the number of iterations 100 until completing setting, and 000, and by image recognition Model is saved；
Further, modified ArcFace loss function calculation formula are as follows:
Wherein, n indicates that the training sample sum of iteration input every time in training process, L indicate that the loss of n sample is average Value, yiIndicate the image category label of i-th of sample,Indicate full articulamentum F2Weight matrix yiColumn and full articulamentum F1Output angle, θjIndicate full articulamentum F2Weight matrix jth column and full articulamentum F1Output angle, s indicate Adaptive cosine coefficient (desirable s=64), C indicate the classification number of total training sample, λ indicate adaptive weighting coefficient (desirable λ= 0.5), m indicates decision edge, is obtained by network training；
In the training process, adaptive cosine coefficient s=64, adaptive weighting coefficient lambda=0.5 are set；
Step (4): image recognition test data set is subjected to image spy using image recognition model obtained in step (3) Sign is extracted, and calculates the cosine similarity P between every two image feature vector, setting image similarity threshold value is T, if similar It spends P and is greater than threshold value T, T=0.8 is set, then judges that two images are same class images, otherwise judges that two images are not same class Image obtains the test result of image recognition model.

Claims (2)

1. a kind of image-recognizing method based on modified ArcFace loss function, it is characterised in that include the following steps:
Step (1): prepare image recognition training dataset, test data set；
Step (2): image recognition network structure of the building based on convolutional neural networks, the image based on convolutional neural networks Identify that network includes convolutional layer, pond layer, full articulamentum, modified ArcFace loss function layer, wherein two convolutional layers and One pond layer constitutes an image recognition minor structure, and image recognition network is by N number of concatenated minor structure, two full articulamentums F1、F2, a modified ArcFace loss function layer composition；
Step (3): image recognition training dataset is input to the image recognition based on convolutional neural networks of step (2) building It is trained in network, the loss function in training process uses modified ArcFace loss function, by constantly carrying out to network Loop iteration training reduces loss function constantly, until completing the number of iterations Q of setting, and image recognition model is protected It deposits；
Step (4): image recognition test data set is subjected to characteristics of image using image recognition model obtained in step (3) and is mentioned It takes, calculates the cosine similarity P between every two image feature vector, setting image similarity threshold value is T, if similarity P Greater than threshold value T, then judge that two images are same class images, otherwise judges that two images are not same class images, obtain image The test result of identification model.
2. a kind of image-recognizing method based on modified ArcFace loss function as described in claim 1, feature exist In modified ArcFace loss function calculation formula in the step (3) are as follows:
Wherein, n indicates that the training sample sum of iteration input every time in training process, L indicate the loss average value of n sample, yi Indicate the image category label of i-th of sample,Indicate full articulamentum F2Weight matrix yiColumn and full articulamentum F1It is defeated Angle out, θjIndicate full articulamentum F2Weight matrix jth column and full articulamentum F1Output angle, s indicates adaptive Cosine coefficient (desirable s=64), C indicate that the classification number of total training sample, λ indicate adaptive weighting coefficient (desirable λ=0.5), m It indicates decision edge, is obtained by network training.
CN201810866142.4A 2018-08-01 2018-08-01 Image identification method based on improved ArcFace loss function Active CN109241995B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810866142.4A CN109241995B (en) 2018-08-01 2018-08-01 Image identification method based on improved ArcFace loss function

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810866142.4A CN109241995B (en) 2018-08-01 2018-08-01 Image identification method based on improved ArcFace loss function

Publications (2)

Publication Number Publication Date
CN109241995A true CN109241995A (en) 2019-01-18
CN109241995B CN109241995B (en) 2021-05-14

Family

ID=65072644

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810866142.4A Active CN109241995B (en) 2018-08-01 2018-08-01 Image identification method based on improved ArcFace loss function

Country Status (1)

Country Link
CN (1) CN109241995B (en)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110659573A (en) * 2019-08-22 2020-01-07 北京捷通华声科技股份有限公司 Face recognition method and device, electronic equipment and storage medium
CN110880018A (en) * 2019-10-29 2020-03-13 北京邮电大学 Convolutional neural network target classification method based on novel loss function
CN111639558A (en) * 2020-05-15 2020-09-08 圣点世纪科技股份有限公司 Finger vein identity verification method based on ArcFace Loss and improved residual error network
CN112200159A (en) * 2020-12-01 2021-01-08 四川圣点世纪科技有限公司 Non-contact palm vein identification method based on improved residual error network
CN112766399A (en) * 2021-01-28 2021-05-07 电子科技大学 Self-adaptive neural network training method for image recognition
CN113255694A (en) * 2021-05-21 2021-08-13 北京百度网讯科技有限公司 Training image feature extraction model and method and device for extracting image features
CN113378833A (en) * 2021-06-25 2021-09-10 北京百度网讯科技有限公司 Image recognition model training method, image recognition device and electronic equipment

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20100082639A1 (en) * 2008-09-30 2010-04-01 Microsoft Corporation Processing maximum likelihood for listwise rankings
CN107145852A (en) * 2017-04-28 2017-09-08 深圳市唯特视科技有限公司 A kind of character recognition method based on homologous cosine losses function

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20100082639A1 (en) * 2008-09-30 2010-04-01 Microsoft Corporation Processing maximum likelihood for listwise rankings
CN107145852A (en) * 2017-04-28 2017-09-08 深圳市唯特视科技有限公司 A kind of character recognition method based on homologous cosine losses function

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
JIANKANG DENG等: "《ArcFace:Additive Angular Margin Loss for Deep Face Recognition》", 《ARXIV》 *
LEI LI等: "《TypicFace:Dynamic Margin Cosine Loss for Deep Face Recognition》", 《PRICAI2018》 *

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110659573A (en) * 2019-08-22 2020-01-07 北京捷通华声科技股份有限公司 Face recognition method and device, electronic equipment and storage medium
CN110659573B (en) * 2019-08-22 2021-03-09 北京捷通华声科技股份有限公司 Face recognition method and device, electronic equipment and storage medium
CN110880018A (en) * 2019-10-29 2020-03-13 北京邮电大学 Convolutional neural network target classification method based on novel loss function
CN110880018B (en) * 2019-10-29 2023-03-14 北京邮电大学 Convolutional neural network target classification method
CN111639558A (en) * 2020-05-15 2020-09-08 圣点世纪科技股份有限公司 Finger vein identity verification method based on ArcFace Loss and improved residual error network
CN111639558B (en) * 2020-05-15 2023-06-20 圣点世纪科技股份有限公司 Finger vein authentication method based on ArcFace Loss and improved residual error network
CN112200159A (en) * 2020-12-01 2021-01-08 四川圣点世纪科技有限公司 Non-contact palm vein identification method based on improved residual error network
CN112766399A (en) * 2021-01-28 2021-05-07 电子科技大学 Self-adaptive neural network training method for image recognition
CN113255694A (en) * 2021-05-21 2021-08-13 北京百度网讯科技有限公司 Training image feature extraction model and method and device for extracting image features
CN113378833A (en) * 2021-06-25 2021-09-10 北京百度网讯科技有限公司 Image recognition model training method, image recognition device and electronic equipment
CN113378833B (en) * 2021-06-25 2023-09-01 北京百度网讯科技有限公司 Image recognition model training method, image recognition device and electronic equipment

Also Published As

Publication number Publication date
CN109241995B (en) 2021-05-14

Similar Documents

Publication Publication Date Title
CN109241995A (en) A kind of image-recognizing method based on modified ArcFace loss function
CN110363122B (en) Cross-domain target detection method based on multi-layer feature alignment
Sun et al. Research on the hand gesture recognition based on deep learning
Zhu et al. High performance vegetable classification from images based on alexnet deep learning model
Zhou et al. BOMSC-Net: Boundary optimization and multi-scale context awareness based building extraction from high-resolution remote sensing imagery
CN105069434B (en) A kind of human action Activity recognition method in video
Li et al. Pedestrian detection based on deep learning model
Pan et al. Intelligent diagnosis of northern corn leaf blight with deep learning model
Wang et al. SSRNet: In-field counting wheat ears using multi-stage convolutional neural network
CN108230330B (en) Method for quickly segmenting highway pavement and positioning camera
CN105975932A (en) Gait recognition and classification method based on time sequence shapelet
CN109034213B (en) Hyperspectral image classification method and system based on correlation entropy principle
Xiao et al. Enhancing multiscale representations with transformer for remote sensing image semantic segmentation
CN108960342A (en) Based on the image similarity calculation method for improving SoftMax loss function
CN105631477A (en) Traffic sign recognition method based on extreme learning machine and self-adaptive lifting
CN104966052A (en) Attributive characteristic representation-based group behavior identification method
CN110072183A (en) Passive type location fingerprint base construction method based on intelligent perception
CN104732247B (en) A kind of human face characteristic positioning method
CN104268553A (en) SAR image target recognition method based on kernel fuzzy Foley-Sammon transformation
CN111695460A (en) Pedestrian re-identification method based on local graph convolution network
Yue et al. Deep learning for hyperspectral data classification through exponential momentum deep convolution neural networks
Wang et al. Weakly supervised object detection based on active learning
Ma et al. Infrared target tracking based on proximal robust principal component analysis method
Zhu et al. Real-time traffic sign detection based on YOLOv2
Zuo et al. An attention-based lightweight residual network for plant disease recognition

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant